Asymmetrical estimator for training encapsulated deep photonic neural networks (2405.18458v3)

Abstract: Photonic neural networks (PNNs) are fast in-propagation and high bandwidth paradigms that aim to popularize reproducible NN acceleration with higher efficiency and lower cost. However, the training of PNN is known to be a challenge, where the device-to-device and system-to-system variations create imperfect knowledge of the PNN. Despite backpropagation (BP)-based training algorithms often being the industry standard for their robustness, generality, and fast gradient convergence for digital training, existing PNN-BP methods rely heavily on the accurate intermediate state extraction for a deep PNN (DPNN). These information accesses truncate the photonic signal propagation, bottlenecking DPNN's operation speed and increasing the system construction cost. Here, we introduce the asymmetrical training (AT) method, tailored for encapsulated DPNNs, where the signal is preserved in the analogue photonic domain for the entire structure. AT's minimum information readout for training bypasses analogue-digital interfaces wherever possible for fast operation and minimum system footprint. AT's error tolerance and generality aim to promote PNN acceleration in a widened operational scenario despite the fabrication variations and imperfect controls. We demonstrated AT for encapsulated DPNN with integrated photonic chips, repeatably enhancing the performance from in-silico BP for different network structures and datasets.